A Validation Study of Patient Interview Data and Pharmacy Records for Antihypertensive, Statin, and Antidepressant Medication Use among Older Women

Transcription

1 American Journal of Epidemiology Copyright 2004 by the Johns Hopkins Bloomberg School of Public Health All rights reserved Vol. 159, No. 3 Printed in U.S.A. DOI: /aje/kwh038 A Validation Study of Patient Interview Data and Pharmacy Records for Antihypertensive, Statin, and Antidepressant Medication Use among Older Women Denise M. Boudreau 1,2, Janet R. Daling 3,4, Kathleen E. Malone 3,4, Jacqueline S. Gardner 2, David K. Blough 2, and Susan R. Heckbert 1,2,4 1 Center for Health Studies, Group Health Cooperative, Seattle, WA. 2 Department of Pharmacy, University of Washington, Seattle, WA. 3 Fred Hutchinson Cancer Research Center, Seattle, WA. 4 Department of Epidemiology, University of Washington, Seattle, WA. Received for publication March 24, 2003; accepted for publication July 31, A validation study evaluated the accuracy of self-reported use of commonly used medications among older women. Within a case-control study of breast cancer, drug information was ascertained by interview. Pharmacy records from 1990 to 1999 were obtained from a Washington State health maintenance organization (66% of subjects) and retail pharmacies (34% of subjects) on a sample of subjects (212 cases, 191 controls) and used as the gold standard. Recall accuracy was assessed for 6-month, 2-year, and 8-year time windows. Sensitivity of antihypertensive use was 92% (95% confidence interval (CI): 85, 96) for cases and controls in the 6-month period and slightly lower for the 2-year (90% (95% CI: 82, 94) and 87% (95% CI: 78, 92)) and 8-year (80% (95% CI: 69, 88) and 79% (95% CI: 68, 88)) periods. For statins, sensitivity was 83% (95% CI: 64, 93) for cases and 93% (95% CI: 69, 99) for controls in the 6-month period, 75% (95% CI: 55, 88) and 86% (95% CI: 60, 96) in the 2-year period, and 67% (95% CI: 42, 85) and 75% (95% CI: 41, 93) in the 8-year period. For self-report of antidepressants, sensitivities ranged from 66% (95% CI: 47, 80) in the 6-month period to 44% (95% CI: 30, 60) in the 8-year period. Specificity was high among all drug classes, ranging from 91% to 100%. Recall did not differ by case-control status. Trivial changes in estimates were observed when health maintenance organization records alone were used as the gold standard. Self-reported use of antihypertensives and statins appears to be relatively accurate among older women. antidepressive agents; antihypertensive agents; hydroxymethylglutaryl-coa reductase inhibitors; interviews; validation studies [publication type]; women Abbreviations: CI, confidence interval; GHC, Group Health Cooperative. Observational studies frequently use self-reported medication data. The quality of these data relies on the ability of subjects to recall information accurately. If medication exposures are not adequately recalled or are recalled differently among cases and controls, the misclassified exposure can lead to erroneous risk and prevalence estimates and results in decreased study power (1). The accuracy of self-reported hormone use has been evaluated in previous studies (2 15), but little work has been done to evaluate the accuracy of self-reported nonhormonal medication use (12, 15 19). In general, recall of hormone replacement therapy and oral contraceptives has been shown to be quite accurate and recall of other medications less accurate. To assess recall accuracy for medications, investigators have commonly used medical records or prescription records as the gold standard. Neither medical records nor pharmacy records are subject to recall bias, but pharmacy records are commonly considered more complete for assessment of medication use (20 23). We used pharmacy records from a health maintenance organization and two retail pharmacy chains to evaluate the accuracy of self-reported medication exposures for three (6-month, 2-year, and 8-year) time windows. The following drug classes were evaluated: anti- Correspondence to Dr. Denise M. Boudreau, Group Health Center for Health Studies, 1730 Minor Avenue, Suite 1600, Seattle, WA ( boudreau.d@ghc.org). 308

2 Validation Study of Interview Data and Pharmacy Records 309 hypertensives, statins (class of lipid-lowering drugs), and antidepressants. MATERIALS AND METHODS Study population Study subjects were a sample of women who participated in a population-based, case-control study of various medications and breast cancer risk (D. M. Boudreau, Group Health Cooperative Center for Health Studies, unpublished manuscript). Cases included 975 women identified from the Surveillance, Epidemiology, and End Results Registry of western Washington State who were diagnosed with primary invasive breast cancer in , whose names appeared on a list of Social Security recipients provided by the Centers for Medicare and Medicaid Services, and who were aged years at the time of diagnosis. The comparison group was 1,007 women without breast cancer randomly selected from the same list of Social Security recipients (24). The reference date was the date of breast cancer diagnosis for cases and a comparable date selected for controls. Subjects were identified for the validation study through informed consent forms that were obtained upon entry into the parent study. The names and locations of pharmacies where women had filled their prescriptions were ascertained, and consent was obtained to access pharmacy records for the past 10 years. Subjects without consent were not available for selection into the validation study (cases = 1.6 percent, controls = 7.3 percent). We selected a subset of cases and controls who reported filling prescription drugs exclusively at the retail pharmacy chains Bartell Drugs and Safeway Incorporated or who were enrollees of the health maintenance organization Group Health Cooperative (GHC) and filled any prescriptions at a GHC pharmacy. Reporting exclusive use of the two pharmacy chains was required because there was no means to validate prescriptions filled at other pharmacies. Bartell Drugs and Safeway Incorporated were chosen because they were among the top 10 most frequently used pharmacies by the case-control study participants, had a large number of branches in the study area, maintained computerized pharmacy records, reported keeping prescription records for more than 2 years, and were willing to participate. Of the 462 women in the parent study who reported using the two retail pharmacies, 158 reported using multiple pharmacies, 65 reported not using the pharmacies at the reference date, and 10 provided too little information for data abstraction. The pharmacy records of only 149 of the 229 remaining eligible subjects (65 percent) who reported exclusive use of either of these retail pharmacy chains were located. There were 288 GHC enrollees identified in the parent study, of whom 94 percent (271/288) were eligible and had pharmacy data available. We limited our study sample to include only women with at least 6 months of pharmacy records before the reference date (n = 403). Data collection In a structured in-person interview, participants provided an inventory of lipid-lowering medications, cardiovascular medications, and antidepressant medications used in the past 20 years. For each course of medication use, the brand name, strength, directions for use, indication(s) for use, and start and stop dates were recorded. Information on current medications was transcribed from the prescription labels on the medication containers. A calendar of life events and a photograph book of lipid-lowering drugs and antihypertensive drugs were used to improve subject recall. Information on the drug name, strength, quantity dispensed, fill date, and directions for use was obtained from pharmacy records for each prescription dispensed at GHC and retail pharmacies. Health maintenance organization enrollment dates were also obtained for GHC enrollees. GHC automated pharmacy records contain data on the prescriptions dispensed at GHC outpatient pharmacies since They are considered to be of high quality and complete, and these records have been used extensively for both primary epidemiologic studies of drug-disease hypotheses and methodological research (25). The completeness and quality of the retail pharmacy records are unknown. Therefore, we estimated our reliability measures of selfreported data using 1) combined data (GHC and retail) as the gold standard comparison and 2) only GHC data as the gold standard comparison. Exposure classification For each time window (6 months, 2 years, and 8 years) prior to the reference date, women were considered true users of a medication class if the pharmacy records indicated that they had filled at least two prescriptions for a drug belonging to that class during the time window. Only GHC enrollees had pharmacy records available for the 8-year period; therefore, reliability measures for this time window were calculated using only GHC data. The antihypertensive medication class included calcium channel blockers, betaadrenergic blockers, angiotensin II receptor antagonists, angiotensin-converting enzyme inhibitors, and diuretics (26). Serotonin reuptake inhibitors, tricyclic antidepressants, and miscellaneous antidepressants were included in the antidepressant drug class (26). are a unique class of drugs used to lower cholesterol (26). We determined the number of days a prescription should last (run-out period) on the basis of the pill quantity dispensed and directions for use. In the event of missing directions for use (42 percent of all dispensings), the average daily dose from all prescriptions with complete information filled for the drug of interest was used to estimate the missing daily dose. If another prescription was filled within 60 days after the end of the run-out period, we assumed that use was continuous during the interval between prescriptions. The duration of exposure to a given drug class was estimated by adding together the total number of days of use within the time windows. Using the interview data, we categorized women as users of a particular therapeutic class of medications if they reported using a drug whose name was included in the

3 310 Boudreau et al. class. The duration of exposure for the interview data was calculated by subtracting the start date from the stop date. As with the pharmacy data, all categorizations of use and estimates of duration were done for each of the three time windows. The majority of medication users had complete data on the start and stop year of use: 89 percent and 94 percent for antihypertensives, 97 percent and 98 percent for statins, and 88 percent and 90 percent for antidepressants. July was imputed for missing start months (60 percent of antihypertensives, 52 percent of statins, and 47 percent of antidepressants), and June was imputed for missing stop months (15 percent of antihypertensives, 14 percent of statins, and 16 percent of antidepressants). January was imputed as the start month when subjects knew the duration of use within a given year but not the exact start and stop month (3 percent of antihypertensives, 15 percent of statins, and 11 percent of antidepressants). If the start year or stop year could not be estimated from the interview, subjects were excluded from the analysis (less than five subjects were excluded from any analysis). Statistical analysis We compared self-reported use with that of combined pharmacy records (GHC and retail) and that of GHC pharmacy records. Using pharmacy records as the gold standard, we calculated the sensitivity, specificity, and respective 95 percent confidence intervals of self-reported drug use for each of the three drug classes (27, 28). The chi-square test for between-group differences was used to detect any differences in sensitivity or specificity between cases and controls. To evaluate the agreement between duration of use as reported by subjects and duration of use calculated from pharmacy records, the intraclass correlation coefficient was calculated using the 1,000-replication naïve bootstrap method (27, 29 32). The 2.5th and 97.5th percentiles of the bootstrap distribution of the 1,000 differences in intraclass correlation coefficients of cases and controls were then estimated (30). The intraclass correlation coefficients of cases and controls were not considered different if the intervals included zero. All reliability measures were estimated for the 6-month, 2- year, and 8-year time windows prior to the reference date. Study subjects without pharmacy records available for the entire time window of interest were excluded from those particular analyses. Analyses were stratified by breast cancer case-control status. RESULTS Cases and controls were comparable in all the characteristics shown in table 1, except that cases were more likely than controls (96 percent vs. 90 percent) to be Caucasian. These characteristics are consistent with participants of the casecontrol study from which our subjects were sampled. Of the 403 eligible women for whom pharmacy records were available for at least 6 months prior to the reference date (66 percent with GHC records, 34 percent with retail pharmacy records), 355 had pharmacy records available for at least 2 years (71 percent with GHC records, 29 percent with retail records), and 222 had pharmacy records available for at least 8 years (100 percent with GHC records). Among the 403 subjects included in the 6-month time window, 212 were breast cancer cases and 191 were controls. The prevalence of any medication use, according to pharmacy records, for each time window and by case-control status is summarized in table 1. Antihypertensive use ranged from 48 percent to 62 percent; statins, from 7 percent to 13 percent; and antidepressants, from 15 percent to 39 percent (table 1). Antihypertensives Of all the women classified as antihypertensive users during the 6 months prior to the reference date according to pharmacy records, 92 percent (95 percent confidence interval (CI): 85, 96) of cases and 92 percent (95 percent CI: 85, 96) of controls reported having used antihypertensives (table 2). The sensitivity of self-reported antihypertensive medication use among cases and controls during the 2 years prior to the reference date was 90 percent (95 percent CI: 82, 94) and 87 percent (95 percent CI: 78, 92). Specificity estimates ranged from 91 percent to 97 percent for the same two time periods. There was little change in the estimates when only GHC records were used as the gold standard (table 3). For the 8-year period, the sensitivity was 80 percent (95 percent CI: 69, 88) among cases and 79 percent (95 percent CI: 68, 88) among controls. Specificity was 100 percent for cases and controls in the 8-year period. Sensitivity and specificity estimates were similar between cases and controls. Intraclass correlation coefficient estimates for selfreported duration of antihypertensive medication use among cases and controls were 0.85 and 0.86 for the 6-month period and 0.89 and 0.90 for the 2-year period (table 4). Intraclass correlation coefficients for duration of antihypertensive medication use during the 6-month and 2-year periods were slightly higher when GHC pharmacy records alone were used as the gold standard (table 5). Intraclass correlation coefficient estimates for the 8-year period were 0.88 for cases and controls, but differences in the median duration of use between self-report and pharmacy records were noted in both groups. There was no difference in intraclass correlation coefficients between cases and controls as indicated by the confidence intervals of the distribution of differences that included zero. The sensitivity of self-reported statin use as compared with combined pharmacy records among cases and controls was 83 percent (95 percent CI: 64, 93) and 93 percent (95 percent CI: 69, 99) for the 6-month period and 75 percent (95 percent CI: 55, 88) and 86 percent (95 percent CI: 60, 96) for the 2-year period (table 2). All specificity estimates were close to 100 percent and did not differ between cases and controls. There was little change in the sensitivity or specificity estimates when only GHC records were used as the gold standard (table 3). Sensitivity and specificity estimates for the 8-year period were 67 percent (95 percent CI: 42, 85) and 100 percent (95 percent CI: 96, 100) in cases and 75

4 Validation Study of Interview Data and Pharmacy Records 311 TABLE 1. Characteristics of validation study subjects, Washington State, Cases Controls Total Patient characteristics No. % No. % No. % Age (years) Race* Caucasian African American Asian/Pacific Islander Other Annual income quartile ($) <20, ,000 34, ,000 49, , Missing Education (years) < > Marital status Single Married/living as married Widowed/divorced/separated Prevalence of medication use In 6 months before reference date Antihypertensives Antidepressants In 2 years before reference date Antihypertensives Antidepressants In 8 years before reference date Antihypertensives Antidepressants * Values significantly different between cases and controls. According to combined (Group Health Cooperative and retail) pharmacy records. Denominator: cases = 212, controls = 191, total = 403. Denominator: cases = 190, controls = 165, total = 355. Denominator: cases = 115, controls = 110, total = 225. percent (95 percent CI: 41, 93) and 99 percent (95 percent CI: 95, 100) in controls. Although controls appeared to recall statin use better than cases did in all the time periods, the sensitivity estimates were not statistically different. Intraclass correlation coefficients for the self-reported duration of statin use among cases and controls were 0.86 and 0.85 for the 6-month period and 0.89 and 0.85 for the 2- year period (table 4). Using only GHC records as the

5 312 Boudreau et al. TABLE 2. Sensitivity and specificity of self-reported drug use compared with pharmacy records for all women (Group Health Cooperative and retail users), by case-control status, Washington State, Cases Controls Therapeutic class Sensitivity (%) 95% CI* Specificity (%) 95% CI Sensitivity (%) 95% CI Specificity (%) 95% CI 6-month period Antihypertensives 92 (103), 85, (105) 85, (92) 85, (99) 86, (24) 64, (188) 95, (14) 69, (177) 95, 99 Antidepressants# 64 (36) 48, (175) 95, (29) 47, (162) 95, 99 2-year period Antihypertensives 90 (99) 82, (88) 89, (91) 78, (74) 91, (24) 55, (166) 97, (14) 60, (151) 95, 100 Antidepressants 56 (43) 41, (146) 95, (31) 41, (134) 93, 99 * CI, confidence interval. Antihypertensives include calcium channel blockers, beta-adrenergic blockers, angiotensin II receptor antagonists, angiotensin-converting enzyme inhibitors, and diuretics. Numbers in parentheses, denominator for the measure. Four subjects were excluded from the 6-month antihypertensive analysis, and three subjects were excluded from the 2-year antihypertensive analysis because of missing data on self-reported use. One subject was excluded from the 6-month and 2-year antidepressant analyses because of missing data on self-reported use. # Antidepressants include tricyclics, serotonin reuptake inhibitors, and miscellaneous (trazodone, buproprion, venlafaxine, and nefaxzodone). comparison group led to trivial changes in the intraclass correlation coefficient estimates (table 5). For the 8-year period, the intraclass correlation coefficient estimates were 0.69 in cases and 0.79 in controls, but differences in the median duration of use between self-report and pharmacy records were noted. There was no difference in intraclass correlation coefficients by case-control status. Antidepressants Among those classified by the combined pharmacy records as antidepressant users during the 6 months prior to the reference date, 64 percent (95 percent CI: 48, 78) of cases and 66 percent (95 percent CI: 47, 80) of controls recalled having used antidepressants during this period (table 2). The sensitivity of self-reported antidepressant medication use TABLE 3. Sensitivity and specificity of self-reported drug use compared with pharmacy records for Group Health Cooperative enrollees only, by case-control status, Washington State, Cases Controls Therapeutic class Sensitivity (%) 95% CI* Specificity (%) 95% CI Sensitivity (%) 95% CI Specificity (%) 95% CI 6-month period Antihypertensives 97 (61), 89, (71) 92, (62) 85, (71) 92, (13) 50, (121) 97, (10) 60, (123) 96, 100 Antidepressants 68 (22) 47, (112) 94, (21) 41, (112) 97, year period Antihypertensivesa 90 (61) 80, (63) 92, (66) 76, (60) 94, (13) 42, (113) 97, (11) 52, (115) 95, 100 Antidepressantsb 61 (28) 42, (98) 96, (24) 35, (102) 96, year period Antihypertensivesa 80 (70) 69, (44) 92, (63) 68, (47) 92, (15) 42, (100) 96, (8) 41, (102) 95, 100 Antidepressantsb 49 (45) 35, (70) 95, (36) 30, (74) 95, 100 * CI, confidence interval. Antihypertensives include calcium channel blockers, beta-adrenergic blockers, angiotensin II receptor antagonists, angiotensin-converting enzyme inhibitors, and diuretics. Numbers in parentheses, denominator for the measure. Two subjects were excluded from the 6-month and 2-year antihypertensives analyses, and one subject was excluded from the 8-year antihypertensive analysis because of missing data on self-reported use. Antidepressants include tricyclics, serotonin reuptake inhibitors, and miscellaneous (trazodone, buproprion, venlafaxine, and nefaxzodone).

6 Validation Study of Interview Data and Pharmacy Records 313 TABLE 4. Comparisons between self-report and pharmacy records for duration of use of various drug classes by all women (Group Health Cooperative and retail users), by case-control status, Washington State, Therapeutic class % of users Self-report Median months of use in users * CI, confidence interval. Antihypertensives include calcium channel blockers, beta-adrenergic blockers, angiotensin II receptor antagonists, angiotensin-converting enzyme inhibitors, and diuretics. Antidepressants include tricyclics, serotonin reuptake inhibitors, and miscellaneous (trazodone, buproprion, venlafaxine, and nefaxzodone). Group Health Cooperative and retail pharmacy records % of users Median months of use in users during the 2 years prior to the reference date was 56 percent (95 percent CI: 41, 70) for cases and 58 percent (95 percent CI: 41, 74) for controls. There was no consistent change in the sensitivity estimates when only GHC records were used as the gold standard for the 6-month and 2-year time periods (table 3). For the 8-year period, the sensitivity was 49 percent (95 percent CI: 35, 63) for cases and 44 percent (95 percent CI: 30, 60) for controls. Specificity estimates tended to approach or equal 100 percent for all time periods, regardless of which pharmacy records were used as the gold standard. All sensitivity and specificity estimates were similar between cases and controls. The intraclass correlation coefficients of self-reported duration of antidepressant medication use compared with those of combined pharmacy records for cases and controls were 0.77 and 0.67 for the 6-month period and 0.81 and 0.74 for the 2-year period prior to the reference date (table 4). Using only GHC records as the comparison group led to trivial changes in the intraclass correlation coefficient estimates and respective 95 percent confidence intervals (table 5). For the 8-year period, the intraclass correlation coefficients were 0.77 for cases and 0.79 for controls. There was no difference in the intraclass correlation coefficients of selfreported duration of antidepressant medication use between cases and controls. DISCUSSION Difference in median Intraclass correlation coefficient 6-month period Antihypertensives Cases , 0.91 Controls , 0.92 Cases , 0.93 Controls , 0.95 Antidepressants Cases , 0.87 Controls , year period Antihypertensives Cases , 0.94 Controls , 0.94 Cases , 0.95 Controls , 0.95 Antidepressants Cases , 0.90 Controls , % CI* In general, the recall accuracy of antihypertensives, antidepressants, and statins was best for the 6-month time window prior to the reference date and decreased slightly for the 2-year and 8-year time windows. The antihypertensive medication class had the highest overall sensitivities, followed by statins, and antidepressants. Our study results suggest some underascertainment and little overascertainment of self-reported medication use. In general, the duration of use was similar between self-report and pharmacy records, but both cases and controls tended to report a longer

7 314 Boudreau et al. TABLE 5. Comparisons between self-report and pharmacy records for duration of use of various drug classes by Group Health Cooperative enrollees only, by case-control status, Washington State, Therapeutic class % of users Self-report Median months of use in users Group Health Cooperative pharmacy records % of users Median months of use in users Difference in median Intraclass correlation coefficient 6-month period Antihypertensives Cases , 0.96 Controls , 0.97 Cases , 0.98 Controls , 0.97 Antidepressants Cases , 0.91 Controls , year period Antihypertensives Cases , 0.95 Controls , 0.98 Cases , 0.97 Controls , 0.95 Antidepressants Cases , 0.90 Controls , year period Antihypertensives Cases , 0.93 Controls , 0.93 Cases , 0.86 Controls , 0.95 Antidepressants Cases , 0.90 Controls , 0.92 * CI, confidence interval. Antihypertensives include calcium channel blockers, beta-adrenergic blockers, angiotensin II receptor antagonists, angiotensin-converting enzyme inhibitors, and diuretics. Antidepressants include tricyclics, serotonin reuptake inhibitors, and miscellaneous (trazodone, buproprion, venlafaxine, and nefaxzodone). 95% CI* duration of use than indicated by the pharmacy records. Recall accuracy of drug use and duration of use did not differ markedly between cases and controls. The recall accuracy reported from similar studies of nonhormonal medications is varied. Most published studies assessed agreement using the kappa statistic, while we used the intraclass correlation coefficient since it does not suffer from the many disadvantages of kappa (33, 34). The intraclass correlation coefficient can be a special case of weighted kappa when the classification categories are equally scaled along one dimension (first category is scored 1, second category is scored 2, etc.) (29). Paganini-Hill and Ross (12) validated self-reported medication use in breast cancer cases and controls, and they found no evidence to indicate that cases recall drug use differently from controls. The study was conducted in women aged years, and it used medical records as the gold standard. The level of agreement between interview and medical

8 Validation Study of Interview Data and Pharmacy Records 315 records was kappa = 0.60 for antihypertensives. For antihypertensive use, the percentage of agreement between medical chart and personal interview was 90 percent among cases and 85 percent among controls. The study did not evaluate antidepressants or statins. The Cardiovascular Health Study compared directed recall and medication inventory for current beta-blocker use (antihypertensive medication class) among adults aged 65 years or older (17). The researchers found a moderate level of agreement with kappa = 0.54 (95 percent CI: 0.51, 0.56) for beta blockers. The Rotterdam elderly study evaluated the agreement between patient interview and pharmacy records for cardiovascular drugs used in the 6 months prior to interview among adults aged 55 or more years (18). Agreement between the two measures varied from poor to almost perfect. However, commonly prescribed antihypertensive agents that were included in our study, such as atenolol (kappa = 0.96), furosemide (kappa = 0.90), and enalapril (kappa = 0.93), were found to have excellent agreement. Kehoe et al. (19) found similar results for recall of antihypertensive medications; self-reported antihypertensive medication use of at least 1 month was reported to have a high sensitivity (88 percent) and specificity (89 percent) when compared with physician report. The accuracy of self-reported exposure to cardiovascular drugs appears to be consistently good across the few published studies (12, 17 19). In addition to supporting these findings, our study extends the validity of antihypertensive medication recall to longer time windows and duration of use and across case-control status. No study to date has evaluated the accuracy of self-reported statin exposure, but we hypothesize that the relatively high sensitivity, specificity, and intraclass correlation coefficient estimates for both antihypertensives and statins are due to similarities in the prescribing of these drug classes. Both drug classes are generally used on a long-term basis, taken daily, and commonly indicated for chronic and often overlapping conditions (26). Unless complications or side effects occur, patients typically remain on the same drug. A photograph book of drugs, including statins and certain classes of antihypertensive drugs, was used in the parent case-control study interview to improve recall. Our estimates of reliability for the drug class of statins are somewhat limited by a low prevalence of drug use. were not regularly prescribed for women until the middle of the 1990s, and the drug class was not commonly included on health maintenance organization formularies for women until that time. Studying the reliability of self-reported statin use is becoming more feasible, as both the prevalence of use and the time available on the market increase. Cotterchio et al. (16) compared the accuracy of selfreported antidepressant medication use with that of physician records for adult female cancer cases (non-hodgkin s lymphoma, breast cancer, and kidney cancer) and controls. The study reported substantial agreement (kappa = 0.60, 95 percent CI: 0.47, 0.74; agreement = 80 percent) for overall lifetime antidepressant medication use ( ever or never ). Moderate agreement (kappa = 0.56, 95 percent CI: 0.32, 0.79) was reported for three categories of lifetime duration of antidepressant medication use, but the level of agreement was somewhat greater for cases than for controls. Our study findings combined with those of Cotterchio et al. (16) indicate only moderate recall accuracy of antidepressant medications, especially as compared with the other medications used for chronic conditions. The degree of misclassification appears to be similar for cases and controls. There are several reasons that may explain why antidepressants were recalled less accurately than antihypertensives and statins in this study. It is not uncommon for patients to misunderstand the indication for an antidepressant prescription. Patients commonly report antidepressant use for symptoms related to depression, such as insomnia or anxiety. In addition, antidepressants have broader indications for use as compared with antihypertensives and statins. It has been suggested that the underreporting of antidepressants and other socially undesirable medications is a result of the reluctance of persons to report such medication use (12, 16). Poor recall may also be related to depression, an indication for the prescription. Finally, unlike statins and antihypertensives, photographs of antidepressant medications were not available for the interview. It is reasonable to speculate that photographs of antidepressants might have improved recall of use. Study limitations should be considered when interpreting our study results. Pharmacy records are not subject to recall bias, but they cannot be considered 100 percent complete and do not address compliance. Patients may use medications differently from what is prescribed, which may explain why self-reported duration is longer than the duration calculated by pharmacy records. To minimize misclassifying nonusers who filled a prescription and did not ingest the drug as users in the pharmacy records, we considered subjects to be users of a drug class only if they filled at least two consecutive prescriptions for drugs belonging to that class. Pharmacy records at GHC and the retail pharmacies cover prescriptions filled at only their pharmacies. Subjects may have filled prescriptions at other pharmacies or obtained drug samples from physicians and, as a result, they may have been misclassified as nonusers according to the pharmacy records. We attempted to minimize this limitation by including only women reporting exclusive use of the retail pharmacies or continuous enrollment in the health maintenance organization, which was confirmed by health maintenance organization enrollment records. As a result, our findings may not be generalizable to all women if women who used multiple pharmacies recall medication use less accurately than do women who used only one pharmacy during the study period. Retail pharmacy records were unavailable for 35 percent of the 229 subjects reporting exclusive use of the two retail pharmacies. Routine purging of electronic pharmacy records after required time periods (2 years in Washington State) is the most likely reason for the missing records, however; it is unknown what influence excluding these subjects has on our findings. We attempted to obtain records on all prescriptions from the retail pharmacies and all prescriptions for the drugs of interest from GHC, but it is possible that prescriptions may have been missed by either source. All of the three drug classes studied contain prescription-only medications in the United States, which, unlike over-the-counter medications, are included in pharmacy records. Only commonly used drugs were included in

9 316 Boudreau et al. the antihypertensive and antidepressant drug categories, and caution should be used in generalizing the study findings to drugs or therapeutic classes not included. Retail pharmacy records are rarely used for research purposes. It is fairly well established that the majority of prescriptions from GHC providers are filled at GHC pharmacies (25), but the use of multiple pharmacies may be more likely among retail pharmacy users. However, we observed little change in the sensitivity, specificity, or intraclass correlation coefficient estimates when GHC pharmacy records alone were used as the gold standard. Our study suggests that older women accurately recall their antihypertensive and statin drug therapy. Recall of antidepressant medication use appeared to be less accurate, although this could be due in part to the less detailed approach used in the study to collect data on antidepressants versus the approach used for the other two drug categories. In general, drug recall does not appear to differ by cancer case-control status and, therefore, studies of the association between drug use and outcome will generally be biased toward the null (1). This bias can be substantial, even with high sensitivity and specificity estimates, if the prevalence of exposure in the nondiseased group is small (35). Care should be taken when generalizing our study findings to other populations and to other medications, but our data should prove useful to other epidemiologists concerned with the accuracy of drug information obtained through self-report. Further research in other populations and for other nonhormonal drug classes is needed. REFERENCES 1. Armstrong BK, White E, Saracci R. Principles of exposure measurement in epidemiology. Vol 21. Oxford, United Kingdom: Oxford University Press, Skegg DCG. Potential for bias in case-control studies of oral contraceptives and breast cancer. Am J Epidemiol 1988;127: Norell SE, Boethius G, Persson I. Oral contraceptive use: interview data versus pharmacy records. Int J Epidemiol 1998;27: Nischan P, Ebeling K, Thomas D, et al. Comparison of recalled and validated oral contraceptive histories. Am J Epidemiol 1993;138: Stolley PD, Tonascia JA, Sartwell PE, et al. Agreement rates between oral contraceptive users and prescribers in relation to drug use histories. Am J Epidemiol 1978;107: Coulter A, Vessey M, McPherson K, et al. The ability of women to recall their oral contraceptive histories. Contraception 1986;33: Rosenberg MJ, Layde PM, Ory HW, et al. Agreement between women s histories of oral contraceptive use and physician records. Int J Epidemiol 1983;12: Adam SA, Sheaves J, Wright NH, et al. A case-control study of the possible association between oral contraceptives and malignant melanoma. Br J Cancer 1981;44: Horowitz RI, Feinstein AR, Stremlau JR. Alternative data sources and discrepant results in case-control studies of estrogens and endometrial cancer. Am J Epidemiol 1980;111: Goodman MT, Nomura AM, Wilkens LR, et al. Agreement between interview information and physician records on the history of menopausal estrogen use. Am J Epidemiol 1990;131: Spengler RF, Clark EA, Woolever CA, et al. Exogenous estrogens and endometrial cancer: a case-control study and assessment of potential biases. Am J Epidemiol 1981;114: Paganini-Hill A, Ross R. Reliability of recall of drug usage and other health-related information. Am J Epidemiol 1982;116: Harlow S, Linet M. Agreement between questionnaire data and medical records. Am J Epidemiol 1989;142: Persson I, Bergkvist L, Adami HO. Reliability of women s histories of climacteric oestrogen treatment assessed by prescription forms. Int J Epidemiol 1987;16: West SL, Savitz DA, Koch G, et al. Recall accuracy for prescription medications: self-report compared with database information. Am J Epidemiol 1995;142: Cotterchio M, Kreiger N, Darlington G, et al. Comparison of self-reported and physician-reported antidepressant medication use. Ann Epidemiol 1999;9: Psaty BM, Lee M, Savage PJ, et al. Assessing the use of medications in the elderly: methods and initial experience in the Cardiovascular Health Study. The Cardiovascular Health Study Collaborative Research Group. J Clin Epidemiol 1992;45: Sjahid SI, Van der Linden PD, Stricker BH. Agreement between the pharmacy medication history and patient interview for cardiovascular drugs: the Rotterdam elderly study. Br J Clin Pharmacol 1998;45: Kehoe R, Wu SY, Leske MC, et al. Comparing self-reported and physician-reported medical history. Am J Epidemiol 1994; 139: West SL, Strom BL, Freundlich B, et al. Completeness of prescription recording in outpatient medical records from a health maintenance organization. J Clin Epidemiol 1994;47: West SL, Strom BL, Poole C. Validity of pharmacoepidemiology drug and diagnosis data. In: Strom B, ed. Pharmacoepidemiology. 2nd ed. Chichester, United Kingdom: John Wiley & Sons, Inc, 2000: Christensen DB, Williams B, Goldberg HI, et al. Comparison of prescription and medical records in reflecting patient antihypertensive drug therapy. Ann Pharmacother 1994;28: Heerdink LR, Leufkens HC, Koppedraaijer C, et al. Information on drug use in the elderly: a comparison of pharmacy, general practitioner, and patient data. Pharm World Sci 1995;17: Apodaca R, Judkins D, Lo A, et al. Sampling from CMS lists. In: Proceedings of the section on survey research methods. Alexandria, VA: American Statistical Association, 1992: Saunders KW, Stergachis A, VonKorff M. Group Health Cooperative of Puget Sound. In: Strom B, ed. Pharmacoepidemiology. 2nd ed. Chichester, United Kingdom: John Wiley & Sons, Inc, 1994: Sewester CS, Dombek CE, Olin BR, et al, eds. Drug facts and comparisons St Louis, MO: Facts & Comparisons, Rothman KJ, Greenland S. Modern epidemiology. 2nd ed. Philadelphia, PA: Lippincott-Raven, Altman DG, Bryant TN, Gardner MJ. Statistics with confidence. 2nd ed. Bristol, United Kingdom: British Medical Journal Books, Fleiss J, Cohen J. The equivalence of weighed kappa and the intraclass correlation coefficient as measures of reliability. Educ Psychol Meas 1973;33: Carpenter J, Bithell J. Bootstrap confidence intervals: when, which, what? A practical guide for medical statisticians. Stat Med 2000;19:

10 Validation Study of Interview Data and Pharmacy Records Morton AP, Dobson AJ. Assessing agreement. Med J Aust 1989;150: Fleiss JL. Statistical methods for rates and proportions. 2nd ed. New York, NY: John Wiley & Sons, Inc, Maclure M, Willett WC. Misinterpretation and misuse of the kappa statistic. Am J Epidemiol 1987;126: Lantz CA, Nebenzahl E. Behavior and interpretation of the kappa statistic: resolution of the two paradoxes. J Clin Epidemiol 1996;49: Copeland KT, Checkoway H, McMichael AJ, et al. Bias due to misclassification in the estimation of relative risk. Am J Epidemiol 1977;105: